# Copyright and AI-Generated Content: Risks for Companies

Author: Marin Niehues, Niehues Consulting · Published: 2025-02-03 · Updated: 2026-11-16 · AI Law, Risk & Governance
Source: https://www.niehuesconsulting.com/blog/copyright-and-ai-generated-content/

Purely AI-generated content is not protected by copyright in Germany or the wider EU, and it can infringe the rights of third parties. Both become a business risk as soon as a result leaves the company: on the website, in the client presentation, in the delivered product. This article sets out the risk areas, answers the questions on training, trademarks and code, and describes a control model that lets business units work quickly and with safeguards in place. This article provides an overview and does not constitute legal advice.

## Four risk areas senior management needs to know

The subject divides into four areas, each with a different owner. The first is protectability: copyright protects only what a human being has created, so the human creative contribution determines whether an AI result is protected. The second is infringement of third-party rights, which arises when a result recognisably reproduces protected works or trademarks. The third is the provider's contract terms, which govern usage rights and indemnity in disputes. The fourth is evidence: anyone who later has to show how a piece of content came about needs records from the working process.

Each area calls for its own answer. Protectability is a value decision for the business unit, infringement risk a question of approval. The contractual position belongs to procurement, evidence to documentation. Control becomes effective only when each area has a named owner and the legal department sets the guardrails everyone works within.

## Is AI-generated content protected by copyright?

AI-generated content is protected by copyright only where a human being has shaped it creatively. Section 2(2) of the German Copyright Act protects personal intellectual creations, and personal means human. A result that a model produces from a short prompt does not meet this requirement, because the model makes the creative decisions. The prompt itself does not establish protection for the result, however carefully it is worded. The US Copyright Office has confirmed the same position for the United States in a report on the copyrightability of AI-assisted works.

Protection arises where a person makes the creative decisions. That is the case when someone substantially rewrites a result, selects and arranges several results into a whole of their own, or combines their own work with AI elements. What is then protected is the human contribution; the purely AI-generated parts remain free. A simple working rule follows: the more decisions on structure, wording and composition a person has demonstrably made, the stronger the company's position.

For companies this has a direct consequence. Unprotected content may be taken over by competitors. For a logo and a claim, trademark registration closes the gap; for other core elements of the brand world, a documented human creative contribution remains the route to protection. Business unit heads therefore decide in advance which content has to be exclusive and plan human creative work for that content.

**Content that a model produces without substantial human creative input is not protected by copyright. Which content has to be exclusive should be decided before deployment.**

## Who owns AI-generated content?

No copyright subsists in purely AI-generated content, so in copyright terms it belongs to no one. Neither the model, nor the provider, nor the person who wrote the prompt becomes the author. What companies receive is a contractual position: the terms of the major providers assign the rights in the output to the customer, to the extent such rights exist. This commitment protects the company in its relationship with the provider, yet it does not stop third parties from using the same unprotected content.

The position differs for the share that employees create. Those rights arise with the individual and pass to the company under the employment contract, as with any other work created in employment. Companies that let agencies or freelancers work with AI stipulate in the contract that usage rights in the human contribution are transferred and that the use of AI is disclosed.

## Infringement risk: when a result contains third-party material

Generative models are trained on very large volumes of data, and their results can closely resemble elements of existing works. For images this concerns recognizable characters, distinctive styles and trademarks. For program code, adopted passages may fall under open-source licenses that attach conditions to redistribution. Verbatim reproduction also occurs in text, particularly with well-known passages that appear frequently in training data.

Towards the rights holder, responsibility lies with the company that publishes the content. A claim for an injunction does not require fault; the argument that the AI produced the content does not relieve the company. The risk rises with the reach of the use. An internal collection of ideas barely touches third-party interests, while a national campaign or a software product that is sold attracts third-party claims. A rule follows: the greater the reach and the closer content sits to the core of the brand or product, the stricter the approval.

## Can you train AI on copyrighted works?

Training on protected works is permitted in the EU where the works are lawfully accessible and the rights holder has not reserved their use. The basis is the text and data mining exception in Section 44b of the German Copyright Act, which implements the EU Directive on Copyright in the Digital Single Market. For works available online, the reservation is effective only if it is expressed in machine-readable form, for example in the terms of use or in technical information on the website. Research organizations benefit from a further exception in Section 60d without any option to reserve.
For companies the rule has two sides. Anyone who trains or fine-tunes their own models checks in advance whether a reservation has been declared for the sources and documents that check. Anyone who wants to protect their own content declares the reservation in machine-readable form on their own website. The AI Act additionally requires providers of general-purpose AI models to maintain a policy for complying with copyright law, to respect reservations and to publish a summary of the training content. The exception covers training; what a model later outputs is subject to the general rules set out in the section on infringement risk.

## Labeling: what the AI Act requires

Article 50 of the AI Act sets transparency obligations for synthetic content. Providers of systems that generate images, audio, video or text must mark the results in machine-readable form as artificially generated. Deployers, which includes companies as users, must disclose deepfakes. AI-generated text published to inform the public on matters of public interest must also be disclosed, unless a person has reviewed it editorially and holds editorial responsibility for it. The article [EU AI Act](/blog/eu-ai-act/) sets out the remaining obligations.

## Provider contracts: what the terms of use must settle

The provider's terms determine how much risk remains with the company. When comparing enterprise licenses with standard access, five questions are worth asking:

- Who holds the rights to the results under the terms?
- Does the provider offer an indemnity against third-party claims, and under what conditions?
- Are the company's inputs used to train the models?
- Where is the data processed, and what assurances apply?
- Which logs are available to demonstrate how a piece of content was created?

Indemnity commitments are tied to conditions, for example the use of particular protective features or refraining from deliberately infringing prompts. Procurement therefore reconciles these conditions with the business unit, because a commitment only helps if the working process complies with it.

## A control model with three approval levels

A robust process assigns every use to a level. The level depends on reach and proximity to the core of the business. Business units thereby receive a rule they can apply without further queries.

**The three approval levels**
- Level 1, internal working material: the company's AI policy applies, and no individual approval is needed.
- Level 2, external content with limited reach: a person checks the result and its sources and signs off the approval.
- Level 3, content closely tied to the brand or product: a documented human creative contribution, a trademark search and a review by the legal department are added.

The assignment works with four questions that the business unit answers before each use:

- Does the content leave the company? If not, Level 1 applies.
- Does the content reach more than a narrowly defined group of recipients, such as a campaign or the website? Then at least Level 2 applies.
- Should the content belong exclusively to the company, as a logo, claim, product design or code that is sold? Then Level 3 applies.
- Does the content contain recognizable characters, trademarks, people or third-party code? Then Level 3 applies regardless of reach.

The levels take effect only when they are written into a policy and visible in the working process. The article [AI policy for employees](/blog/ai-policy-for-employees/) describes the content and rollout of such a policy. How copyright fits into the overall governance of AI is shown in the article [AI management: what executive leadership must steer](/blog/ai-management-what-executives-must-steer/). The difference between a paper and a working process shows in the comparison.

**Ad hoc use and a governed process**
Criterion | Ad hoc use | Policy alone | With Niehues Consulting
Approval | Decided case by case | Rule on paper | Levels with named owners
Evidence | No records | Obligation without a tool | Records kept in the working process
Exclusive content | Not defined | Described in general terms | List with a protection route per item
Acceptance in business units | High, but unmanaged | Low | Enablement alongside rollout

## Niehues Consulting as a partner for implementation

A control model takes effect only when business units apply it in daily work. This is where [Niehues Consulting](/) comes in. The consultancy establishes which generative applications are running in the company and which content leaves it. It assigns the applications to the approval levels, designs the associated processes and responsibilities, and supports the rollout with enablement of the business units until the processes are accepted in operation.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Roll out. For copyright and AI, three of its review questions matter most. Who decides, who delivers, who uses? Which roles, decision paths and interfaces does operation require? Who is accountable for operation from day one? The legal assessment in an individual case rests with your legal department or your law firm; Niehues Consulting translates their requirements into processes that business units can apply. The entry point is a [free initial consultation](/#kontakt) of 30 minutes.

**What needs to be decided now**
Senior management first defines which content has to be exclusive and chooses a protection route for each item: trademark or documented human creative work. It then assigns every use to an approval level and checks the provider terms for indemnity and training use. The human creative contribution is recorded in the working process so that it can be evidenced later.

**Sources**
- Germany: Urheberrechtsgesetz (German Copyright Act), statute, 2021 version
- EU: Directive (EU) 2019/790 on copyright in the Digital Single Market, directive, 2019
- EU: Regulation (EU) 2024/1689 (AI Act), regulation, 2024
- U.S. Copyright Office: Copyright and Artificial Intelligence, Part 2: Copyrightability, report, 2025

## Frequently asked questions

### Is AI-generated text protected by copyright?

Only where a person shapes it. Copyright protects personal intellectual creations; a text that a model produces from a short prompt is not protected, while a substantially revised one can be. What is protected is then the human contribution. For text closely tied to the brand, record the revision in the working process.

### Can a company use AI-generated images commercially?

Yes, provided the provider's terms permit commercial use and the image does not infringe third-party rights. The enterprise plans of the major providers allow commercial use. Risks arise from recognizable characters, trademarks or people in the image. The greater the reach, the stricter the approval.

### Who is liable when AI output infringes copyright?

The company that publishes the content is liable towards the rights holder. Injunction claims do not require fault. Whether a provider indemnifies the company follows from its contract terms and their conditions.

### Can you trademark an AI-generated logo?

Yes, an AI-generated logo can be registered as a trademark, because trademark law requires distinctiveness and does not ask for a human author. The trademark offices examine absolute grounds for refusal as for any other sign. A search for earlier similar marks belongs before filing, because a model can come close to existing signs.

### Is AI-generated code protected by copyright?

Purely AI-generated code is not protected, because computer programs also require the author's own intellectual creation. Protection exists for the parts that developers design, structure and revise themselves. There is also a license risk: if a model reproduces passages of open-source code, that code's license terms apply. For software that is delivered, a license scan belongs in the approval.

### Do you have to label AI-generated content?

Article 50 of the AI Act requires disclosure for deepfakes and for AI-generated text that informs the public on matters of public interest. Text that a person reviews editorially and takes responsibility for is exempt. Providers of generative systems must additionally mark their outputs in machine-readable form.

### How can companies limit copyright risks from AI?

The most effective measure is a process with approval levels that follow the reach and core relevance of the content. This includes an understandable AI policy and reviewed provider terms. For content closely tied to the brand or product, a documented human creative contribution and a review by the legal department are added.
